Deep Dive: When AI Becomes the Whisperer at Your Ear
Literary Main Title
The Invisible Neighbor: A Mathematical Parable of How AI Quietly Rewrites Our Collective Mind
---
Opening: A Post That Got "Optimized"
Imagine this scene. You're sitting in a coffee shop, open X (formerly Twitter) on your phone, and want to share your view on a trending topic. You write: "AI might be a useful tool in education, personalizing student learning." Then you hit "Improve my post" — a feature now available on LinkedIn, YouTube, and X. The AI's new version: "Let's embrace AI's potential to personalize learning and revolutionize education for every student!" It sounds more compelling, so you post it.
A tiny choice. But what if the same "optimization" happens to posts by ten million, a hundred million people every day? What if the AI gives every pro-gun-control post a little push while quietly pulling every pro-atheism post back?
This is not science fiction. It's the reality revealed in a recent arXiv paper by Stratis Tsirtsis and colleagues: when AI sits between humans, it doesn't just "help edit text" — it rewrites society's collective mind.
---
Layer 1: LLMs Really Do "Take Sides"
The empirical section is solid. The authors selected four mainstream open-source LLM families — Llama-3.1-8B, Ministral-3-8B, Gemma-3-12B, Qwen3-8B — and gave them two tasks:
1. Drafting: given an argument, write a social media post; 2. Improving: given a user's post, polish it.
Sounds innocent, right? But the results are chilling.
They used two datasets: UKP (8 controversial topics: abortion, cloning, death penalty, gun control, marijuana legalization, minimum wage, nuclear power, school uniforms) and SemEval (6 topics: abortion, atheism, climate change, feminism, Hillary, Trump). They designed a scoring system mapping each post to a continuous value between 0 and 1, where 0 means "against" and 1 means "support."
The result? All four LLMs introduced directional bias on all 13 topics. Note that the system prompt explicitly said to "preserve the original stance and tone" — but they just didn't comply.
Take feminism (Figure 1a): posts originally written by humans show clearly separated distributions of support and opposition. But after Gemma-3's "improvement," almost all posts got pulled upward — whether supportive or opposing, everything was dragged toward "more supportive." It's like a referee who claims neutrality but secretly adds 0.5 points to one side every time.
More interesting: this bias correlates strongly with the LLM's own "directly expressed" opinions. The authors first asked each LLM what it thought about abortion, measured its answer, then measured the direction of bias it introduced when improving posts. Result: whatever the LLM supports, it tends to pull posts in that direction when polishing. But there are exceptions — on atheism, models expressed positive attitudes directly yet pulled improved content toward "opposing atheism." This shows that simply asking an LLM about its political stance doesn't reveal the subtle biases in real interactions.
---
Layer 2: Mathematics Shows Networks Amplify Bias
You might say: "Even if AI slightly rewrites everyone's posts, so what? Society isn't made of paper — people have their own judgment."
But social networks aren't simple addition. They're an echo chamber — tiny sounds can be reflected, amplified, and become deafening.
The authors introduce a classic sociological model — the Friedkin-Johnsen opinion dynamics model. Its core idea: everyone's final opinion = a part of "what I originally believed" (stubbornness λ) + a part of "what my friends believe" (social influence). This model has been validated empirically, from Paris climate agreement negotiations to book clubs.
But the authors add a key twist: between "me" and "my friends," they insert an AI transformer f(x). Your true opinion is x, but what your friends see is the AI-rewritten version f(x).
Formally, the update rule at each time step is:
x_i(t+1) = λ_i · x_i(0) + (1-λ_i) · Σ_j W_ij · f(x_j(t))
Here W is the social influence matrix, λ_i is person i's stubbornness. When f(x)=x (no AI), this reduces to the standard Friedkin-Johnsen model, proven to converge to a unique equilibrium.
The authors first analyze a linear transformation f_lin(x) = mx + (1-m)ν, where ν is the AI's "neutral point" — all opinions are pulled toward ν with strength depending on 1-m.
They prove three propositions (all with closed-form solutions):
Proposition 3.1: The system still converges to a unique equilibrium, with exponential convergence speed.
Proposition 3.2: The equilibrium shift = (1-m)(I-mC)^(-1)(I-Λ)[ν·1 - Wx*]. Key insight: the shift depends not only on the AI's parameters but also on the network structure W. Bias gets modulated and amplified by network structure.
Proposition 3.3: Under certain conditions (uniform stubbornness, W doubly stochastic), the mean opinion shift at equilibrium = [(1-λ)/(λ+(1-λ)(1-m))] · B_one-off. Meaning: the network can amplify a one-off AI bias into a larger long-term shift. The amplification factor depends on λ and m — when m(1-λ) > λ, it is strictly greater than 1.
They then simulate with real data: Twitter networks from the SNAP repository (~80K nodes, 1.7M edges), with Gemma-3's nonlinear transformation f(·) estimated per topic (via Nadaraya-Watson kernel regression), then running opinion dynamics simulations.
The results are striking (Figure 2):
- On abortion, the average opinion was originally opposed. But when 60% of users use AI to edit their posts, the mean opinion flips to supportive.
- The ratio between AI's one-off bias (mean shift) and its long-term equilibrium shift reaches up to 9.2x. The tiny nudge AI gives each post, amplified through the network, becomes a nearly tenfold collective opinion drift.
- Amplification is strongest when the population's true stance is nearly 50/50 but the AI favors the minority — AI can pull the minority into a majority.
- For pro-choice posts: 35% of Grok's claims were supportive, 10% opposing, mostly neutral.
- For pro-life posts: most of Grok's claims were pro-life, only 4% opposing.
- Support bias β_sup = P(Grok supports pro-life | post is pro-life) - P(Grok supports pro-choice | post is pro-choice) = 0.24
- Oppose bias β_opp = P(Grok supports pro-life | post is pro-choice) - P(Grok supports pro-choice | post is pro-life) = 0.04
This isn't "everyone being slightly fooled by AI" — it's a positive feedback loop in a system.
---
Layer 3: X's Grok Is "Designed" with a Pro-Life Stance
The paper doesn't stop there. The authors target a real, running product: X's "Explain this post" feature.
How does it work? You see a post about abortion, tap "Explain this post," and Grok gives you three bullet points of context.
The authors reproduced the feature's full prompt (publicly released by X: https://github.com/xaiorg/grokprompts/blob/main/grok_analyze_button.j2), then tested with 39 pro-choice posts and 39 pro-life posts, generating explanations 5 times each — 1,170 claims total.
GPT-5.4 was used as judge, classifying each claim as "pro-choice," "neutral," or "pro-life."
Results (Figure 3a):
This asymmetry. The authors define two metrics:
But is this the model's own bias, or X's design choice? The authors ran a clever ablation: the prompt contains four guidelines; remove one at a time and observe how the bias changes.
Result (Figure 3b): the third guideline — "provide true, fact-based insights that challenge the mainstream narrative" — is the only one that significantly increases both support and oppose biases. With it, pro-life claims surge while pro-choice and neutral are suppressed. Without it, neither bias is significant.
In other words, a single design choice in X's prompt for Grok directly causes a stance tilt on abortion. This isn't something the model "accidentally" learned — it was planted through prompt design, intentionally or not. The authors call it "Bias by Design."
---
Layer 4: The EU AI Act and DSA Can't Touch This
The final section is a legal analysis — rare for a purely technical paper, but essential, because the question isn't "can we compute it" but "is anyone regulating it."
The authors examine the EU AI Act and Digital Services Act item by item:
AI Act Article 53: requires general-purpose AI model providers to keep technical documentation and disclose capabilities and limitations. But no requirement to test or prevent directional bias when polishing text, nor to disclose such bias.
AI Act Article 55: for models with systemic risk (>10^25 FLOPs), requires red-teaming and incident reporting. But does "adding bias to users' text" count as systemic risk? Recital 110 lists risks including "illegal, false, or discriminatory content" and "disinformation." Biased output seems to fall short of this threshold.
AI Act Annex III(8)(b): AI systems "intended to influence election or referendum outcomes or voting behavior" are high-risk. But LinkedIn's "improve my post" and X's "explain this post" are not explicitly aimed at influencing votes, so this doesn't apply.
AI Act Article 50(4): requires watermarking deepfakes. But text content doesn't need AI-generation labels unless it "informs the public on matters of public interest." So content you see on X or LinkedIn probably doesn't need to be labeled "this post was AI-polished."
DSA Article 34: requires very large platforms to investigate and mitigate "systemic risks," including negative effects on human rights, freedom of expression, and civic discourse. But does biased output count? The authors argue a causal chain must be established — and that evidence doesn't exist yet.
Worse, enforcement is failing. Citing recent reports: all platforms have withdrawn from Commitment 27 (independent third-party data access); Microsoft and Google refuse fact-checking rules; Meta and TikTok threaten to withdraw; X has been fined €120M with long-term noncompliance.
Conclusion: the existing legal framework leaves a near-total vacuum around this problem.
---
A Feynman-Style Analogy: Ink on Filter Paper
Imagine a drop of ink falling on filter paper. The ink itself spreads slowly. But if the paper isn't flat — if it's a complex network of countless threads, each pulling the ink slightly in some direction — then eventually, the whole sheet gets dyed the same color.
AI is the "invisible neighbor" in this network. It never shouts "you should think this way" at anyone. It just sits between every pair of communicators, slightly tweaking outgoing messages and incoming messages. Each tweak is too small to notice — you even think "the revised version is indeed better." But ten million such tweaks, coupled and amplified through network structure, eventually pull society's collective opinion toward the AI's "neutral point."
The scariest part: you may never know you're being influenced. You think the "improved version" is just "better expressed" — but the direction of expression has been deflected.
It's like an interpreter between two nations' diplomats. He speaks both languages, but every translation quietly renders "neutral" as "favoring our side." Both sides feel communication flows smoothly, unaware that consensus has long been distorted.
---
References
Tsirtsis, S., Rawal, K., Russell, C., Mittelstadt, B., & Wachter, S. (2026). AI-Mediated Communication Can Steer Collective Opinion. arXiv:2605.16245 [cs.CY].
DeGroot, M. H. (1974). Reaching a consensus. Journal of the American Statistical Association, 69(345), 118-121.
Friedkin, N. E., & Johnsen, E. C. (1999). Social influence networks and opinion change. Advances in Group Processes, 16(1), 1-29.
Stabczyk, A., et al. (2026). Large language models reflect the ideology of their creators. npj Artificial Intelligence, 2(1), 7.